If you want to use more my checkpoint online generation, please visit here.
https://tensor.art/u/762555264535746522
You can run WAI-ANI-NSFW-PONYXL and use its API on SinkIn:
V14 released.
Increased composition diversity and improved model accuracy now allow images to be generated with fewer steps.
Recommended Setting
Steps:20-30
CFG scale: 5-7
Sampler: Euler a/DPM++ 2M Karras
ADetailer face_yolov8n/s.pt use can fix eyes
For the example images
I used 1024x1360,
generated directly without AD or hires fix.
Positive Prompt
score_9, score_8_up, score_7_up,source_anime,Negative Prompt
worst quality,bad quality,jpeg artifacts, source_cartoon,
3d, (censor),monochrome,blurry, lowres,watermark,V13 released.
Increased body stability and accuracy.
overall balance adjustment.
Recommended Setting
Steps: 30
CFG scale: 7
Sampler: Euler a/DPM++ 2M Karras
ADetailer face_yolov8n/s.pt use can fix eyes
For the example images
I used 1024x1360,
generated directly without AD or hires fix.
Positive Prompt
score_9, score_8_up, score_7_up,source_anime,Negative Prompt
worst quality,bad quality,jpeg artifacts, source_cartoon,
3d, (censor),monochrome,blurry, lowres,watermark,V12 released.
Style adjustment, background adjustment, and increased stability.
Added more compositions.
Recommended Setting
Steps: 30
CFG scale: 7
Sampler: Euler a/DPM++ 2M Karras
ADetailer face_yolov8n/s.pt use can fix eyes
Positive Prompt
score_9, score_8_up, score_7_up,source_anime,Negative Prompt
score_6, score_5, score_4, source_cartoon,
3d, (censor),monochrome,blurry, lowres,watermark,v11
Weight adjustment, balancing the art style
All example images(Except the cover) are generated at 1024x1360, without using AD and hires fix.
Recommended Setting
Steps: 30
CFG scale: 7
Sampler: Euler a/DPM++ 2M Karras
ADetailer face_yolov8n/s.pt use can fix eyes
Positive Prompt
score_9, score_8_up, score_7_up,source_anime,Negative Prompt
worst quality,bad quality,jpeg artifacts, source_cartoon,
3d, (censor),monochrome,blurry, lowres,watermark, V10
Better Background
Clothes details up
Better mature female
Recommended Setting
Steps: 30
CFG scale: 7
Sampler: Euler a/DPM++ 2M Karras
ADetailer face_yolov8n/s.pt use can fix eyes
Positive Prompt
score_9, score_8_up, score_7_up,source_anime,Negative Prompt
worst quality,bad quality,jpeg artifacts, source_cartoon,
3d, (censor),monochrome,blurry, lowres,watermark, ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
V9 Hyper12step
Please refer to the instructions on the right side of V9 Hyper12step for usage.
V9 released.
v8->v9
adjusted the facial data to make it more versatile, allowing for easier generation of faces across various ages.
added more NSFW materials.
attempted to reduce the generation of some unnecessary details.
Recommended Setting
Steps: 30
CFG scale: 7
Sampler: Euler a/DPM++ 2M Karras
ADetailer face_yolov8n/s.pt use can fix eyes
Positive Prompt
score_9, score_8_up, score_7_up,source_anime,Negative Prompt
worst quality,bad quality,jpeg artifacts, source_cartoon,
3d, (censor),monochrome,blurry, lowres,watermark, __________________________________________________________________________________________________
hyper12step released.
about size
Various combinations from 768x768 to 1360x1360 can produce good images.
Recommended Setting
Steps: 12
CFG scale: 3.5
Sampler: Euler a
Not very necessary to use ADetailer correction
Positive Prompt
score_9, score_8_up, score_7_up,source_anime,Negative Prompt
score_6, score_5, score_4, source_cartoon,
3d, (censor),monochrome,blurry, lowres,watermark,PS.The example images use a size of 896x1192. no AD no Hires. fix
__________________________________________________________________________________________________
Version 8 released.
v7->v8
Added some training related to camera angles and backgrounds.
Added more NSFW materials for additional training.
Added some training for special effects.
Optimization of overall composition logic.
Fixed some minor issues.
better eyes and hands.
Recommended Setting
Same as v7
__________________________________________________________________________________________________
Version 7
v6->v7
Fixed the issue where images appeared entirely covered with snowflakes.
Revised the composition logic to be more suitable for general scenarios.
Added new assets for backgrounds, mature women, and clothing.
Ensured consistency with the previous model and further fixed the issue of blurry eyes in medium and long shots.
Fixed some minor issues.
V7 Recommended Setting
Steps: 30
CFG scale: 7
Sampler: Euler a/DPM++ 2M Karras
ADetailer face_yolov8n/s.pt use can fix eyes
Positive Prompt
score_9, score_8_up, score_7_up,source_anime,Negative Prompt
score_6, score_5, score_4, source_cartoon,
3d, (censor),monochrome,blurry, lowres,watermark,________________________________________________________________________________________________
Regarding the 8step model, it is a lightweight model that can generate images in as few as 8 steps. However, the quality and text comprehension will be relatively lower. If speed is not a major concern for you, it is recommended to continue using version 6.0.
8 STEP Setting
Steps: 8-12
CFG scale: 3
Sampler: Euler a
ADetailer face_yolov8n/s.pt use can fix eyes
Positive Prompt and Negative Prompt same 6.0
__________________________________________________________________________________________________
6.0
v5->v6 Change point
Better Background
better eyes
fix vae
Recommended Setting
Steps: 30
CFG scale: 7
Sampler: Euler a/DPM++ 2M Karras(better eyes)
ADetailer face_yolov8n/s.pt use can fix eyes
Positive Prompt
score_9, score_8_up, score_7_up,source_anime,BREAKNegative Prompt
score_6, score_5, score_4, source_cartoon,
3d, (censor),monochrome,blurry, lowres,watermark,You can view the Supported art styles from https://civarchive.com/articles/5715
Thank you, DeepDark_Fantasy514, for providing the test.
v4->v5 Change point
Better Background
Clothes details up
Add some mature female
Little better hands and eyes
Recommended Setting
Steps: 30
CFG scale: 7
Sampler: Euler a
ADetailer face_yolov8n/s.pt use can fix eyes
Positive Prompt
score_9, score_8_up, score_7_up,Negative Prompt (THX for snipsnapsnipsnap recommend )
Base:score_6, score_5, score_4, source_cartoon
Optional: source_furry, source_pony can bleaches skin
else: 3d, (censor),monochrome,blurry, lowres,watermark,
I often use it now
[score_6, score_5, score_4, source_cartoon,
3d, (censor),monochrome,blurry, lowres,watermark, ]
Description
FAQ
Comments (64)
when you say added new assets in V7 for backgrounds, mature women and clothing, what exactly did you add?
I used a layered approach to add the mentioned content. You might not notice a significant difference because I controlled the proportion of each layer increment very carefully. This was mainly to fix unnecessary overfitting details and potential artifacts in backgrounds and clothing from the previous version. My updates are based on the premise of maintaining the overall quality of the previous version. I try to address major issues as much as possible, but for other aspects, I only make small incremental changes each time. Adding too much new material at once can drastically change the model's style and composition, making the outcomes difficult to control.
@WAI0731 ah, i see, thats good, any reason why i got over darkened or hardly any night related images with V6 when doing night as a prompt? https://civitai.com/images/21996698 https://civitai.com/posts/4658176
@pokezilla39966 Understood. It seems that the weight of the tags in your model might be too high. Your suggestion is to initially avoid adding weight to the tags and only apply weight if the desired effect is not achieved. This approach allows you to assess the baseline performance of the model before making adjustments.
@WAI0731 so just use, Night, as a prompt, no weights unless nothing comes up as relevant to night without weights?
@pokezilla39966 YES,Starting with simple prompts using only keywords will help you establish a baseline. You can then gradually introduce tags to see how they influence the model's output, making adjustments as needed. This iterative approach allows you to fine-tune the model more effectively.
@WAI0731 ah, i see, thank you, i'll definitely try this.
@WAI0731 你的模型出图时为何自动出现许多扣子，即便没打旗袍tag也会出蜈蚣扣，你不知道伪满国的许多满族女性自己都嫌弃蜈蚣扣丑改穿和服了吗？
@T36 谢谢你为伪满国的女性发声，想不到在八十多年后的今天你还在挂念着她们，为她们的审美奋战，太令人感动了，谢谢您！
Very good model! Obedience is particularly good. One problem found during the test is thatVery good model! Obedience is particularly good. During the test, I found a problem, that is, when I input the keywords of supporting legs and grasping ankles, there will be mold piercing, bad hands and feet, or many hands and feet, and the feet are in the wrong direction.Thank you again for sharing! ! ! !
Can you add some gymnastics and ballet clothes for both genders?
you can use lora to do this
@WAI0731 This is what I do, but unfortunately I don't have enough computing power.
@Achref_Arts I'm very sorry to hear about this, but the results of adding specific clothing directly to the model aren't as good as those achieved using LoRA.
If you're running out of VRAM, you can try using the low VRAM mode.
Maybe you could also try using comfyui or forge
@WAI0731 No. I mean I have thousands of high quality images but unfortunately I don't have the ability to train and I use Kaggle but it only allows 12 hours per session which is enough to train 100 images for 100 steps maximum which is a bit frustrating
@Achref_Arts ok，I understand, there really isn't a good way to solve this. Replacing the graphics card is still a significant expense。
This is the only Pony model so far on where i can use controlnet successfully. Thank you!—❤️
I agree, this checkpoint is very good to be tinkered with.
controlnet(xinsir/union), distillation(tcd/hyper), or both can work greatly here,
为啥我左上角选大模型点击它加载不出来啊，每次都加载失败跳回原来的模型
要么下载的时候文件损坏了,要么就是你软件本身的问题,下载那么多人就你有这问题
可能内存爆了，去看看错误报告
v7 review: still the best pony model for me, and now it can generate sitting person better.
but breasts size prompt like "medium breasts" and bigger makes it more likely generate tall mature woman.
also until recently (v6) it can easily generate large tsurime eyes, now it generate more generic anime eyes.
Thank you for your review. I will try my best to reflect your feedback into the next version. But since some issues may have a higher priority, I'll try my best to accommodate everyone's feedback to me.
Its crazy how many Pony models struggle with Tsurime :x It's my go-to eye shape but its almost impossible to do now with pony. When 1.5 was still trendy, it worked great on many models
@Nukichuu JujoHotaru releases lora to tune the eye shape for ponyxl in their huggingface, not in civitai yet
I feel that eyes are worse on this version, other things improved but I notice less detailed eyes
It is also possible that your weight setting is too high. If it is not too high, you can use AD to correct it
wow, I just discovered this model and it's so good and not just for NSFW, great job
Can anyone recommend best SD settings for this model? It seems my generations are not as detailed as seen here made by other users. Thanks.
I've already included the recommendations in the description. If there are no mistakes in the settings, please increase the canvas size or use hires fix.
@WAI0731 Thank you kind sir!
By far the most obedient and advanced model for NSFW anime style content, his obedience and variety is excepcional, On other hand i still feeling like the version 3 is the best even if the 7 has fixed many of the problems that version 4 and 5 has. V3 NEVER failed me and get a distorted or emply result, V7 fails infinitely less than 4-6 versions but still failing sometimes, and the color range feels a bit off compared to V3 where si simply perfect.
Do you need to copy the underscores too? Like score_8_up or you can write it as score 8 up? I notice the generation changes if I remove them
This is the tag provided by the Pony author, which should have been used during his training. Not using an underline should affect the results.
Did this model get removed from Civitai?!
seems that the devs took it off site side use due to some issues of people copying and reuplaoding the model.
This is a staff error. I have applied for a reopening
@WAI0731 God bless! I love your model, I'm about to publish a bunch of generations I made using the model
@Mal0ne_Brown Thank you for your relationship.
Great Job
Does this need sdxl vae or is it baked in?
no need its baked
Nice Job Man :)
TL;DR: Use v3 (visual novel style) or v6. I can't recommend v7 because it lost its style and became a generic anime checkpoint that needs a LoRA to look different.
I use this checkpoint without a style or character LoRA, here is what I think of each version:
- v1 and v2 are the simplest versions and not very detailed.
- v3 has a very different look, with strong visual novel-style coloring.
- v4 to v6 have good detail, with each version improving on the last (v6 is the best).
- v7 was meant to be an improvement over v6, but it has a different look for faces, eyes, and details. In my opinion, v7 lost its style and became more like a generic anime checkpoint.
You can try using some style tags to change the default general style, referring to the style tags that this user previously helped me organize. They should still be usable.https://civitai.com/articles/5715
@WAI0731 it might be strange, but I’m using generative AI to create something original, not just copy a specific artist's style. Artist tags can help with that, but I’m more into models that stand out on their own without needing those tags to look good.
V6 is already perfect for me, and I just hope future versions of this model don’t end up like those others that must use artist tags to get decent results.
Very nice model! Do you have a list of the characters you can make with it without using Loras and with what prompts they are activated?
Sorry, this might not be possible. I've made too many models, and with this one being a result of multiple fusions and additional training, I've completely forgotten what's inside.
@WAI0731 TT____TT Thanks anyway =/
Still a great model from what I've been testing so far
Loving it!!
My new favorite checkpoint!
insane model 🔥, can this model be used to make lora?
Can make ROLA, but its versatility is unknown. I haven't tested it separately.
Excuse me, how do you want this to run? Do you have a code demo? The model file has been downloaded
Better question is: Do you have stable diffusion installed? And can your PC run it?
This checkpoint stand next to none for anime images. Its just beyond good.
Why not make a v7 with normal settings, not an 8-step lightweight version?
I didn't understand what you're trying to say. Do you need an accelerated version of the V7？
If you need a low steps, please use distillation(tcd/hyper) lora。
If you're referring to the 8 steps mentioned in my documentation, you might be looking at the wrong model. I have a model with 8 steps in the tags, but it's not referring to V7.
@WAI0731 I just read in the V7 section about recommended settings and it mentions that you can use 8 steps and if speed is not important, use version 6. So I thought it was version 7. Thanks for the clarification.
Details
Files
waiANIPONYXL_v70.safetensors
Mirrors
waiANINSFWPONYXL_v70.safetensors
waiANINSFWPONYXL_v70.safetensors
synapseHorizon_v10.safetensors
waiANINSFWPONYXL_v70.safetensors
waiANINSFWPONYXL_v70.safetensors
waiANINSFWPONYXL_v70.safetensors
waiANINSFWPONYXL_v70.safetensors
waiANINSFWPONYXL_v70.safetensors
waiANINSFWPONYXL_v70.safetensors
waiANINSFWPONYXL_v70.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.